Skip to main content

Run your first task

Take one small change from a goal to a checked result in your own repository. AgentD Local orchestration is free. Workers use your coding client and model account, so model usage and tool subscriptions are billed separately.

Before you start​

Use Linux with systemd user services, macOS, or Windows with Git for Windows. You also need Git 2.25 or newer and an installed Claude Code or Codex client. Sign in to the client or configure your model API access. Your repository needs at least one commit. The GitHub CLI is needed for GitHub pull-request actions. See requirements for setup details.

Install and activate​

  1. Open the AgentD page in LibraOS Console and run its install command for your platform. Installing needs no account: the installer is public and verifies release checksums.

  2. Running needs an activation code. On the same page, redeem your beta invitation with the email address it was sent to, and keep the code ready for setup.

  3. In your terminal, check the installed binary and open your repository:

    export PATH="$HOME/.local/bin:$PATH"
    agentd version
    cd /path/to/your-repo

  4. Set your identity and initialize AgentD:

    export AGENTD_USER=YOUR_NAME
    agentd init
    agentd doctor

    Replace YOUR_NAME with your own name, conventionally your GitHub login. Keep AGENTD_USER in your shell profile for new sessions. On Windows, skip the export PATH line and open a new terminal after installing; in PowerShell, set $env:AGENTD_USER = "YOUR_NAME" instead, and init saves it for new terminals. Setup asks for your activation code at a hidden prompt and for EULA acceptance; the agreement is at libraos.com/eula, and agentd eula prints the text your binary carries. Enter the code in your own terminal, never in a chat or command argument.

    Init creates the local state and services, connects installed coding clients, registers the repository and adds you as its owner. Follow agentd doctor until no required check fails. Connection details cover provider keys, client configuration and team identities.

Define the task and its check​

Choose a change you can verify independently. For a JavaScript repository with a lockfile and an existing test script, a first task could be:

  • Goal: add keyboard support to the command menu.
  • Acceptance: Escape closes the menu and focus returns to its trigger.
  • Check: the project's tests cover both keyboard behaviors.

Run this from a clean checkout, after choosing the check command your project actually uses:

agentd run "Add keyboard support to the command menu: Escape closes it and focus returns to its trigger. Add tests for both behaviors." --check "npm ci && npm test" --budget 30m

This starts a real worker using your model account. Use --worker codex to select Codex. Dependency installation needs network access; see checks and dependencies before using a restricted worker.

Follow the result​

The terminal result names the task, branch, commit and check result. If you detach with Ctrl-C, use agentd handoff show ID, replacing ID with the task ID printed by AgentD.

In the dashboard, Team shows running workers, Issues holds the change and its evidence, and Inbox brings back questions, failures and decisions that need you.

Confirm the check ran against the change and that its evidence addresses the acceptance criteria. A failed or skipped check needs attention. For an independent assessment, give a separate validator the goal, criteria and result; see validation.

Deliver when ready​

Publishing and merging are separate authorized actions. Open a draft pull request from the task when the result is ready; an owner can merge through the dashboard subject to GitHub permissions and branch protection. Run, verify and deliver explains the controls.

For your next task, plan work in Chat, connect a different model through Gateway, or add optional Runtime knowledge services.